Automatic cognate identification with gap-weighted string subsequences.

Taraka RamaNAACL 2015

TL;DR

This paper introduces a novel approach to automatically identify cognates (words with a common ancestral origin) by using "gap-weighted string subsequences." This method, inspired by string kernels, outperforms traditional word similarity measures and existing methods by effectively capturing phonetic similarities and accounting for sound changes over time.

What Problem Does This Paper Address?

Identifying cognates is crucial for historical linguistics and has applications in NLP tasks like language reconstruction and comparative lexicography. While traditional methods rely on string similarity or phonetic similarity, they often struggle with words that have diverged significantly from their common ancestor due to sound changes (like sound loss, gain, or assimilation). This paper tackles the challenge of accurately identifying cognates, especially those that are not obviously similar in form, by proposing a more robust feature representation.

Methods

The core contribution of this paper is the development of gap-weighted string subsequence features. These features are inspired by string kernels, which are known for their ability to capture similarities between strings without strict positional constraints.

Here's a breakdown:

  • String Subsequences: Instead of looking at contiguous character sequences (like n-grams), this method considers subsequences, where characters can be "skipped" (gaps).
  • Gap-Weighting (λ): A parameter λ is introduced to exponentially penalize longer gaps. This means subsequences with fewer skipped characters are weighted more heavily. As λ approaches 1, the feature essentially counts the frequency of a subsequence; as λ approaches 0, it's constrained to substrings.
  • Phonetic Integration: The authors propose integrating phonetic features into the vector representation. This is achieved by mapping phonetic symbols to broader classes (like Consonant/Vowel) or potentially to an intermediary phonetic alphabet.
  • Feature Vector: The final feature vector for a word pair is a combination of these subsequence features, normalized to a unit vector.
  • Classifier: A Support Vector Machine (SVM) with L2-regularization and L2-loss is used as the classifier.

Datasets

The primary dataset used is the publicly available Indo-European dataset from Dyen et al. (1992).

  • Dyen et al. (1992): This dataset contains 16,520 lexical items for 200 concepts across 84 language varieties, transcribed in a broad romanized phonetic alphabet.
  • Dataset Creation: Word pairs were extracted for each concept. A pair was labeled as "cognate" (positive) if it shared an identical Cognate Class Number (CCN), and "non-cognate" (negative) otherwise. The authors created a large dataset of 619,635 word pairs, with 145,145 cognates.
  • Other Datasets Explored: The authors also considered the ASJP (Automated Swadesh-list Project) and Ringe et al. (2002) datasets. However, they found ASJP to have limited cognacy judgments and no Indo-European family data, while Ringe et al.'s dataset was in orthographic format and had missing items, making them less suitable for their experiments.

Key Insights and Ideas

  • Subsequences are More Robust than N-grams/Edit Distance: The gap-weighted subsequence approach effectively handles words that have diverged significantly due to sound changes, which traditional edit distance or n-gram measures might miss. For example, the edit distance between "hound" and "hund" is 1, but the difference between "wheel" and "chakra" is substantial, yet they are cognates. Subsequences can bridge these gaps.
  • Phonetic Features Enhance Cognate Identification: Integrating phonetic information (even at a broad level like C/V) into the feature vectors is shown to improve performance, as it captures sound correspondences more directly.
  • The Importance of Gap Penalization (λ): The parameter λ plays a crucial role in balancing the importance of exact matches versus allowing for sound changes. Finding the optimal λ is key to maximizing performance.
  • Non-Unigram Models Outperform: The experiments suggest that considering subsequences longer than single characters (non-unigram models) generally leads to better performance.

Key Results

The paper demonstrates that their proposed gap-weighted subsequence features outperform the baseline system (Hauer and Kondrak, 2011 - HK) which uses traditional word-similarity features.

  • Cross-Validation Accuracy: The best accuracy achieved with their system was at λ = 0.8 and p = 3, outperforming the HK system's accuracy of 82.61%.
  • F1-Score: The best performing subsequence system (λ = 0.7, p = 2) achieved an F1-score of 0.5, which is higher than the HK system's F1-score of 0.46.
  • Robustness: The system showed robustness across different random splits of the dataset.
  • Overall: The subsequence features, particularly non-unigram models, consistently performed better than the HK-based classifier in both cross-validation and concept-based experiments.

Limitations and Future Work

  • Dataset Size and Scope: The authors acknowledge that a larger dataset of cognacy judgments for other language families would be beneficial.
  • Phonetic Transcription: The current dataset is not transcribed in a detailed IPA format, limiting the full potential of phonetic feature integration.
  • Future Directions:
    • Creating larger datasets for more language families.
  • Utilizing richer phonetic transcriptions (e.g., IPA).
  • Integrating articulatory phonetic features.
  • Testing various combinations of feature vectors.

Summary updated

BibTeX
@inproceedings{rama2015automatic,
  author = {Taraka Rama},
  title = {{Automatic cognate identification with gap-weighted string subsequences.}},
  booktitle = {Conference of the North American Chapter of the Association for Computational Linguistics},
  year = {2015},
}
Figure